Subject: Re: Ouch, ouch ouch. NetBSD/SPARC 1.0 on IPC gets SPECint92 rating of 2.
To: Paul Kranenburg <pk@cs.few.eur.nl>
From: Daniel Carosone <danielce@ee.mu.oz.au>
List: port-sparc
Date: 02/04/1995 11:08:51
Paul Kranenburg writes:
 > > 
 > > netbsd4me:1:28 [/tmp] % /usr/bin/time sh -c "echo '2^8192' | bc > /dev/null"
 > >        93.26 real        70.64 user         0.61 sys
 > > 
 > something must be wrong:
 > 
 > NetBSD chroom 1.0A NetBSD 1.0A (VOLVOX) #56: Wed Feb  1 22:38:45 MET 1995
 >     pk@chroom:/net/baloe/bsd/sys/arch/sparc/compile/VOLVOX sparc
 > % /usr/bin/time sh -c "echo '2^8192' | bc > /dev/null"
 >         2.77 real         2.55 user         0.16 sys
 > 
 > (a Sun ELC)

pk is being rather modest here. Something is wrong, and he knows
probably better than anyone just what it is. Some months ago, this
very issue of why dynamically-linked binaries ran so slowly on sparc
came up. After quite a bit of intricate work, he discovered and
removed a feature of the way mmap and shared libs were interacting
with the cpu cache.

He maintains that the new code is not quite fixed for all cases, but
after the fixes went in my sparc system got *VERY NOTICABLY* faster.

I'd recommend you try -current. If you want more info, at least some
of what transpired should be in the port-sparc archives.

--
Dan.